The University of Applied Sciences and Arts Bielefeld (HSBI) presents the development status of the method for replacing non-linear algebraic loops with machine learning surrogates at the annual MODPROD (Center for Model-Based Cyber-Physical Product Development) workshop in Linköping, Sweden.
When solving complex equation-based models some equations need to be evaluated simultaneously and are called strong components or loops. If these equations are non-linear a typical solving algorithm is the Newton-Raphson method. These systems of non-linear equations can take up a significant portion of the computational cost when solving the ordinary differential equation system.
HSBI presents a workflow to automatically detect non-linear equation systems in Modelica models and replace them with machine learning surrogates. The development concentrates on efficient ways to generate large amounts of training data from an Model-Exchange Functional Mock-up Unit (FMU) and how to replace the non-linear solver with the trained neural network.
Workshop: MODPROD Annual Workshop